Visual scene recognition is an indispensable part of automatic localization and navigation. In the same scene, the appearance and viewpoint may be changed greatly, which is the largest challenge for some advanced unmanned systems,e.g. robot,vehicle and UAV,etc., to identify scenes where they have visited. Traditional methods have been subjected to hand-made feature-based paradigms for a long time, mainly relying on the prior knowledge of the designer, and are not sufficiently robust to extreme changing scenes. In this paper, we cope with scene recognition with automatically learning the representation of features from big image samples. Firstly, we propose a novel approach for scene recognition via training a slight-weight convolutional neural network (CNN) that overall has less complex and more efficient network architecture, and is trainable in the manner of end-to-end. The proposed approach uses the deep-leaning features of self-selection combining with light CNN process to perform high semantic understanding of visual scenes. Secondly, we propose to employ a salient region-based technology to extract the local feature representation of a specific scene region directly from the convolution layer based on self-selection mechanism, and each layer performs a linear operation with end-to-end manner. Furthermore, we also utilize probability statistics to calculate the total similarity of several regions in one scene to other regions, and finally rank the similarity scores to select the correct scene. We have conducted a lot of experiments to evaluate the results of performance by comparing four methods (namely, our proposed and other three well known and advanced methods). Experimental results show that the proposed method is more robust and accurate than other three well-known methods in extremely harsh environments (e. g. weak light and strong blur).

视觉场景识别是自动定位和导航不可或缺的一部分。在同一个场景中，外观和视点可能会发生很大的变化，这对于一些先进的无人系统，如机器人、车辆和无人机等，识别它们访问过的场景是最大的挑战。传统方法长期以来受制于手工制作的基于特征的范式，主要依赖于设计者的先验知识，对极端变化的场景不够鲁棒。在本文中，我们通过从大图像样本中自动学习特征表示来处理场景识别。首先，我们提出了一种通过训练轻量级卷积神经网络 (CNN) 进行场景识别的新方法，该网络总体上具有更简单、更高效的网络架构，并且可以以端到端的方式进行训练。所提出的方法使用自选择的深度学习特征结合轻型 CNN 过程来对视觉场景进行高度语义理解。其次，我们建议采用基于显着区域的技术，基于自选择机制直接从卷积层中提取特定场景区域的局部特征表示，每一层以端到端的方式进行线性运算。此外，我们还利用概率统计来计算一个场景中多个区域与其他区域的总相似度，最后对相似度分数进行排序以选择正确的场景。我们进行了大量实验，通过比较四种方法（即，我们提出的和其他三种众所周知的先进方法）。实验结果表明，在极端恶劣的环境（如弱光和强模糊）下，所提出的方法比其他三种众所周知的方法更鲁棒和准确。